Martin Fowler says LLM-written prose carries a voice you cannot edit out

Martin Fowler published a short bliki post on 29 October. The title is "LLM Voice". The argument is that readers can feel an LLM-written draft. They may not say why. Editing afterwards does not remove the voice. The post reached 156 points on Hacker News.

Fowler's argument has three parts. The first is that LLM-written prose has a voice. The second is that the voice is detectable by a reader, not just by a tool. The third is that editing afterwards does not fully remove it. "I feel I'm reading something that comes from an alien culture", he wrote.

Why writing is the thinking

The core line in the post is a working writer's reason, not a stylistic one. Fowler points to writing as the step where a thought becomes clear to its author. "Often it's only by writing something down that I can really get my thoughts into coherence", he wrote. Hand that step to a model, Fowler argues, and you have handed over the understanding behind it. The post was not written as a style guide. It was written to protect the thinking.

Fowler names two exceptions he accepts. The first is accessibility. Jim Highsmith, a long-time friend, uses an LLM to overcome the effects of Parkinson's disease, Fowler writes. The second is translation for a non-native English speaker. The LLM lets a non-English-first writer reach an English-speaking audience. Fowler does not accept the general case: using an LLM to speed up a draft.

What Fowler now asks of contributors

Fowler also stated the disclosure rule for his own site. "I did not use AI or LLMs in any way while writing this post", he wrote in the piece itself. Contributors must declare any AI use in the acknowledgments of their posts on martinfowler.com. The rule treats AI help the same way an academic journal treats a co-author.

The piece is not about engineering. It is about writing as a step in thinking. That is why it reached a developer audience on Hacker News. A commenter named "jonahx" argued on the thread that cutting the writing step also cuts the author's reasoning. That argument is Fowler's point.

What a team publishing content does with this

A team using an LLM in the draft of its blog or product manual has a test to run. Read the last five published pieces aloud. Ask whether a reader at a buying desk would finish them. The piece Fowler put up has a signal that an LLM draft rarely holds. That signal is a thought the writer worked out while writing.

The alternative is to edit. Fowler says editing does not catch the voice fully. A writer who treats the LLM output as a starting point still ends with a flat piece. The honest answer, in Fowler's post, is to use the model for research and lookups. Write the sentences yourself.

Fowler closes the post by saying that writing is the way to think. There are better ways to use AI than at that step, he writes. His own practice backs it up. The site is a 25-year bliki with his byline on every entry. The new rule asks every contributor to say whether a model helped.

Source

LLM Voice, Martin Fowler, 29 October. Discussion: Hacker News thread, 156 points.